ddtrace 0.45.0 → 0.50.0

data/lib/ddtrace/profiling/collectors/stack.rb

@@ -0,0 +1,262 @@
+require 'ddtrace/profiling/backtrace_location'
+require 'ddtrace/profiling/events/stack'
+require 'ddtrace/utils/only_once'
+require 'ddtrace/utils/time'
+require 'ddtrace/worker'
+require 'ddtrace/workers/polling'
+
+module Datadog
+  module Profiling
+    module Collectors
+      # Collects stack trace samples from Ruby threads for both CPU-time (if available) and wall-clock.
+      # Runs on its own background thread.
+      #
+      # rubocop:disable Metrics/ClassLength
+      class Stack < Worker
+        include Workers::Polling
+
+        DEFAULT_MAX_TIME_USAGE_PCT = 2.0
+        MIN_INTERVAL = 0.01
+        THREAD_LAST_CPU_TIME_KEY = :datadog_profiler_last_cpu_time
+
+        attr_reader \
+          :recorder,
+          :max_frames,
+          :ignore_thread,
+          :max_time_usage_pct,
+          :thread_api
+
+        def initialize(
+          recorder,
+          max_frames:,
+          ignore_thread: nil,
+          max_time_usage_pct: DEFAULT_MAX_TIME_USAGE_PCT,
+          thread_api: Thread,
+          fork_policy: Workers::Async::Thread::FORK_POLICY_RESTART, # Restart in forks by default
+          interval: MIN_INTERVAL,
+          enabled: true
+        )
+          @recorder = recorder
+          @max_frames = max_frames
+          @ignore_thread = ignore_thread
+          @max_time_usage_pct = max_time_usage_pct
+          @thread_api = thread_api
+
+          # Workers::Async::Thread settings
+          self.fork_policy = fork_policy
+
+          # Workers::IntervalLoop settings
+          self.loop_base_interval = interval
+
+          # Workers::Polling settings
+          self.enabled = enabled
+
+          @warn_about_missing_cpu_time_instrumentation_only_once = Datadog::Utils::OnlyOnce.new
+
+          # Cache this proc, since it's pretty expensive to keep recreating it
+          @build_backtrace_location = method(:build_backtrace_location).to_proc
+          # Cache this buffer, since it's pretty expensive to keep accessing it
+          @stack_sample_event_recorder = recorder[Events::StackSample]
+        end
+
+        def start
+          @last_wall_time = Datadog::Utils::Time.get_time
+          reset_cpu_time_tracking
+          perform
+        end
+
+        def perform
+          collect_and_wait
+        end
+
+        def loop_back_off?
+          false
+        end
+
+        def collect_and_wait
+          run_time = Datadog::Utils::Time.measure do
+            collect_events
+          end
+
+          # Update wait time to throttle profiling
+          self.loop_wait_time = compute_wait_time(run_time)
+        end
+
+        def collect_events
+          events = []
+
+          # Compute wall time interval
+          current_wall_time = Datadog::Utils::Time.get_time
+          last_wall_time = if instance_variable_defined?(:@last_wall_time)
+                             @last_wall_time
+                           else
+                             current_wall_time
+                           end
+
+          wall_time_interval_ns = ((current_wall_time - last_wall_time).round(9) * 1e9).to_i
+          @last_wall_time = current_wall_time
+
+          # Collect backtraces from each thread
+          thread_api.list.each do |thread|
+            next unless thread.alive?
+            next if ignore_thread.is_a?(Proc) && ignore_thread.call(thread)
+
+            event = collect_thread_event(thread, wall_time_interval_ns)
+            events << event unless event.nil?
+          end
+
+          # Send events to recorder
+          recorder.push(events) unless events.empty?
+
+          events
+        end
+
+        def collect_thread_event(thread, wall_time_interval_ns)
+          locations = thread.backtrace_locations
+          return if locations.nil?
+
+          # Get actual stack size then trim the stack
+          stack_size = locations.length
+          locations = locations[0..(max_frames - 1)]
+
+          # Convert backtrace locations into structs
+          locations = convert_backtrace_locations(locations)
+
+          thread_id = thread.respond_to?(:pthread_thread_id) ? thread.pthread_thread_id : thread.object_id
+          trace_id, span_id = get_trace_identifiers(thread)
+          cpu_time = get_cpu_time_interval!(thread)
+
+          Events::StackSample.new(
+            nil,
+            locations,
+            stack_size,
+            thread_id,
+            trace_id,
+            span_id,
+            cpu_time,
+            wall_time_interval_ns
+          )
+        end
+
+        def get_cpu_time_interval!(thread)
+          # Return if we can't get the current CPU time
+          unless thread.respond_to?(:cpu_time_instrumentation_installed?) && thread.cpu_time_instrumentation_installed?
+            warn_about_missing_cpu_time_instrumentation(thread)
+            return
+          end
+
+          current_cpu_time_ns = thread.cpu_time(:nanosecond)
+
+          # NOTE: This can still be nil even when all of the checks above passed because of a race: there's a bit of
+          # initialization that needs to be done by the thread itself, and it's possible for us to try to sample
+          # *before* the thread had time to finish the initialization
+          return unless current_cpu_time_ns
+
+          last_cpu_time_ns = (thread[THREAD_LAST_CPU_TIME_KEY] || current_cpu_time_ns)
+          interval = current_cpu_time_ns - last_cpu_time_ns
+
+          # Update CPU time for thread
+          thread[THREAD_LAST_CPU_TIME_KEY] = current_cpu_time_ns
+
+          # Return interval
+          interval
+        end
+
+        def get_trace_identifiers(thread)
+          return unless thread.is_a?(::Thread)
+          return unless Datadog.respond_to?(:tracer) && Datadog.tracer.respond_to?(:active_correlation)
+
+          identifier = Datadog.tracer.active_correlation(thread)
+          [identifier.trace_id, identifier.span_id]
+        end
+
+        def compute_wait_time(used_time)
+          # We took used_time to get the last sample.
+          #
+          # What we're computing here is -- if used_time corresponds to max_time_usage_pct of the time we should
+          # spend working, how much is (100% - max_time_usage_pct) of the time?
+          #
+          # For instance, if we took 10ms to sample, and max_time_usage_pct is 1%, then the other 99% is 990ms, which
+          # means we need to sleep for 990ms to guarantee that we don't spend more than 1% of the time working.
+          used_time_ns = used_time * 1e9
+          interval = (used_time_ns / (max_time_usage_pct / 100.0)) - used_time_ns
+          [interval / 1e9, MIN_INTERVAL].max
+        end
+
+        # Convert backtrace locations into structs
+        # Re-use old backtrace location objects if they already exist in the buffer
+        def convert_backtrace_locations(locations)
+          locations.collect do |location|
+            # Re-use existing BacktraceLocation if identical copy, otherwise build a new one.
+            @stack_sample_event_recorder.cache(:backtrace_locations).fetch(
+              # Function name
+              location.base_label,
+              # Line number
+              location.lineno,
+              # Filename
+              location.path,
+              # Build function
+              &@build_backtrace_location
+            )
+          end
+        end
+
+        def build_backtrace_location(_id, base_label, lineno, path)
+          string_table = @stack_sample_event_recorder.string_table
+
+          Profiling::BacktraceLocation.new(
+            string_table.fetch_string(base_label),
+            lineno,
+            string_table.fetch_string(path)
+          )
+        end
+
+        private
+
+        def warn_about_missing_cpu_time_instrumentation(thread)
+          @warn_about_missing_cpu_time_instrumentation_only_once.run do
+            # Is the profiler thread instrumented? If it is, then we know instrumentation is available, but seems to be
+            # missing on this thread we just found.
+            #
+            # As far as we know, it can be missing due to one the following:
+            #
+            # a) The thread was started before we installed our instrumentation.
+            #    In this case, the fix is to make sure ddtrace gets loaded before any other parts of the application.
+            #
+            # b) The thread was started using the Ruby native APIs (e.g. from a C extension such as ffi).
+            #    Known cases right now that trigger this are the ethon/typhoeus gems.
+            #    We currently have no solution for this case; these threads will always be missing our CPU instrumentation.
+            #
+            # c) The thread was started with `Thread.start`/`Thread.fork` and hasn't yet enabled the instrumentation.
+            #    When threads are started using these APIs, there's a small time window during which the thread has started
+            #    but our code to apply the instrumentation hasn't run yet; in these cases it's just a matter of allowing
+            #    it to run and our instrumentation to be applied.
+            #
+            if thread_api.current.respond_to?(:cpu_time) && thread_api.current.cpu_time
+              Datadog.logger.debug(
+                "Thread ('#{thread}') is missing profiling instrumentation; other threads should be unaffected"
+              )
+            end
+          end
+        end
+
+        # If the profiler is started for a while, stopped and then restarted OR whenever the process forks, we need to
+        # clean up the per-thread cpu time counters we keep, so that the first sample after starting doesn't end up with:
+        #
+        # a) negative time: At least on my test docker container, and on the reliability environment, after the process
+        #    forks, the clock reference changes and (old cpu time - new cpu time) can be < 0
+        #
+        # b) large amount of time: if the profiler was started, then stopped for some amount of time, and then
+        #    restarted, we don't want the first sample to be "blamed" for multiple minutes of CPU time
+        #
+        # By resetting the last cpu time seen, we start with a clean slate every time we start the stack collector.
+        def reset_cpu_time_tracking
+          thread_api.list.each do |thread|
+            thread[THREAD_LAST_CPU_TIME_KEY] = nil if thread[THREAD_LAST_CPU_TIME_KEY]
+          end
+        end
+      end
+      # rubocop:enable Metrics/ClassLength
+    end
+  end
+end

data/lib/ddtrace/profiling/encoding/profile.rb

@@ -0,0 +1,37 @@
+require 'set'
+require 'time'
+
+require 'ddtrace/profiling/flush'
+require 'ddtrace/profiling/pprof/template'
+
+module Datadog
+  module Profiling
+    module Encoding
+      module Profile
+        # Encodes gathered data into the pprof format
+        module Protobuf
+          module_function
+
+          def encode(flush)
+            return unless flush
+
+            # Create a pprof template from the list of event types
+            event_classes = flush.event_groups.collect(&:event_class).uniq
+            template = Pprof::Template.for_event_classes(event_classes)
+
+            # Add all events to the pprof
+            flush.event_groups.each { |event_group| template.add_events!(event_group.event_class, event_group.events) }
+
+            Datadog.logger.debug do
+              "Encoding profile covering #{flush.start.iso8601} to #{flush.finish.iso8601}, " \
+              "events: #{flush.event_count} (#{template.debug_statistics})"
+            end
+
+            # Build the profile and encode it
+            template.to_pprof
+          end
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/event.rb

@@ -0,0 +1,13 @@
+module Datadog
+  module Profiling
+    # Describes a sample of some data obtained from the runtime.
+    class Event
+      attr_reader \
+        :timestamp
+
+      def initialize(timestamp = nil)
+        @timestamp = timestamp || Time.now.utc.to_f
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/events/stack.rb

@@ -0,0 +1,102 @@
+require 'ddtrace/profiling/event'
+
+module Datadog
+  module Profiling
+    module Events
+      # Describes a stack profiling event
+      class Stack < Event
+        attr_reader \
+          :frames,
+          :hash,
+          :span_id,
+          :thread_id,
+          :total_frame_count,
+          :trace_id
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          trace_id,
+          span_id
+        )
+          super(timestamp)
+
+          @frames = frames
+          @total_frame_count = total_frame_count
+          @thread_id = thread_id
+          @trace_id = trace_id
+          @span_id = span_id
+
+          @hash = [
+            thread_id,
+            trace_id,
+            span_id,
+            [
+              frames.collect(&:hash),
+              total_frame_count
+            ]
+          ].hash
+        end
+      end
+
+      # Describes a stack sample
+      class StackSample < Stack
+        attr_reader \
+          :cpu_time_interval_ns,
+          :wall_time_interval_ns
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          trace_id,
+          span_id,
+          cpu_time_interval_ns,
+          wall_time_interval_ns
+        )
+          super(
+            timestamp,
+            frames,
+            total_frame_count,
+            thread_id,
+            trace_id,
+            span_id
+          )
+
+          @cpu_time_interval_ns = cpu_time_interval_ns
+          @wall_time_interval_ns = wall_time_interval_ns
+        end
+      end
+
+      # Describes a stack sample with exception
+      class StackExceptionSample < Stack
+        attr_reader \
+          :exception
+
+        def initialize(
+          timestamp,
+          frames,
+          total_frame_count,
+          thread_id,
+          trace_id,
+          span_id,
+          exception
+        )
+          super(
+            timestamp,
+            frames,
+            total_frame_count,
+            thread_id,
+            trace_id,
+            span_id
+          )
+
+          @exception = exception
+        end
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/exporter.rb

@@ -0,0 +1,23 @@
+require 'ddtrace/profiling/transport/io/client'
+
+module Datadog
+  module Profiling
+    # Writes profiling data to a given transport
+    class Exporter
+      attr_reader \
+        :transport
+
+      def initialize(transport)
+        unless transport.is_a?(Profiling::Transport::Client)
+          raise ArgumentError, 'Unsupported transport for profiling exporter.'
+        end
+
+        @transport = transport
+      end
+
+      def export(flush)
+        transport.send_profiling_flush(flush)
+      end
+    end
+  end
+end

data/lib/ddtrace/profiling/ext/cpu.rb

@@ -0,0 +1,52 @@
+module Datadog
+  module Profiling
+    module Ext
+      # Monkey patches Ruby's `Thread` with our `Ext::CThread` to enable CPU-time profiling
+      module CPU
+        # We cannot apply our CPU extension if a broken rollbar is around because that can cause customer apps to fail
+        # with a SystemStackError: stack level too deep.
+        #
+        # This occurs whenever our extensions to Thread are applied BEFORE rollbar applies its own. This happens
+        # because a loop forms: our extension tries to call Thread#initialize, but it's intercepted by rollbar, which
+        # then tries to call the original Thread#initialize as well, but instead alls our extension, leading to stack
+        # exhaustion.
+        #
+        # See https://github.com/rollbar/rollbar-gem/pull/1018 for more details on the issue
+        ROLLBAR_INCOMPATIBLE_VERSIONS = Gem::Requirement.new('<= 3.1.1')
+
+        def self.supported?
+          unsupported_reason.nil?
+        end
+
+        def self.apply!
+          return false unless supported?
+
+          # Applying CThread to Thread will ensure any new threads
+          # will provide a thread/clock ID for CPU timing.
+          require 'ddtrace/profiling/ext/cthread'
+          ::Thread.prepend(Profiling::Ext::CThread)
+          ::Thread.singleton_class.prepend(Datadog::Profiling::Ext::WrapThreadStartFork)
+        end
+
+        def self.unsupported_reason
+          # NOTE: Only the first matching reason is returned, so try to keep a nice order on reasons -- e.g. tell users
+          # first that they can't use this on macOS before telling them that they have the wrong ffi version
+
+          if RUBY_ENGINE == 'jruby'
+            'JRuby is not supported'
+          elsif RUBY_PLATFORM.include?('darwin')
+            'Feature requires Linux; macOS is not supported'
+          elsif RUBY_PLATFORM =~ /(mswin|mingw)/
+            'Feature requires Linux; Windows is not supported'
+          elsif !RUBY_PLATFORM.include?('linux')
+            "Feature requires Linux; #{RUBY_PLATFORM} is not supported"
+          elsif Gem::Specification.find_all_by_name('rollbar', ROLLBAR_INCOMPATIBLE_VERSIONS).any?
+            'You have an incompatible rollbar gem version installed; ensure that you have rollbar >= 3.1.2 by ' \
+            "adding `gem 'rollbar', '>= 3.1.2'` to your Gemfile or gems.rb file. " \
+            'See https://github.com/rollbar/rollbar-gem/pull/1018 for details.'
+          end
+        end
+      end
+    end
+  end
+end